Could machines decide who to attack? US and Russia weaken UN AI rules

United Nations: The United States and Russia have forced the removal of key safeguards from a proposed United Nations framework on lethal autonomous weapons, including a provision requiring human review of targets selected by artificial intelligence, according to a report by The Washington Post.

The changes were made during negotiations held in Geneva earlier this month under the Convention on Certain Conventional Weapons. The discussions marked the UN's most advanced effort to establish international rules for weapons capable of identifying, selecting and engaging targets with limited or no human intervention.

US and Russian diplomats spent nearly 15 hours revising the document on the final day of negotiations, the Post reported, citing three people familiar with the closed-door discussions and documents reviewed by the newspaper.

The two delegations removed provisions requiring autonomous weapons to operate in a “predictable” and “reliable” manner. They also deleted a clause calling for ethical considerations in the use of such systems.

A provision requiring human operators to review military targets identified by AI before an attack was also removed, the report said.

The negotiations were held behind closed doors after UN cameras were switched off and civil society observers were asked to leave. The US and Russian delegations each included around 10 legal experts, nearly twice the number fielded by many other countries, the Post reported.

The US State Department, Russian Foreign Ministry and the United Nations did not respond to requests for comment from the newspaper.

Human Rights Watch flags weakened safeguards

Human Rights Watch separately said the United States and Russia had weakened provisions that had received broad support from other governments.

The final document removed references to the design and development of autonomous weapons and narrowed the scope of the legal framework from international law generally to international humanitarian law.

It also dropped references to explainability and traceability, safeguards intended to help establish how an AI system arrived at a targeting decision and determine responsibility when something goes wrong.

The revised document, however, retains language stating that human control and judgement are necessary to ensure compliance with international law. It also includes restrictions on systems that cannot operate within the rules governing armed conflict.

Verity Coyle of Human Rights Watch warned that weaker safeguards could allow machines to make life-and-death decisions without effective human control.

She said the potential consequences included greater civilian harm, reduced accountability and the faster expansion of automated warfare.

UN framework could lead to binding treaty

The Geneva process currently involves non-binding recommendations. However, the document could serve as the basis for negotiations on the world's first legally binding treaty regulating lethal autonomous weapons.

Decisions under the Convention on Certain Conventional Weapons are generally reached by consensus, giving major military powers considerable influence and allowing individual countries to block stronger provisions supported by a majority.

Governments are expected to meet again in Geneva in November to decide whether the three-year deliberations should move towards formal negotiations on a new international instrument.

AI's growing role in warfare

The debate has gained urgency as artificial intelligence becomes increasingly integrated into military planning and operations.

The Post previously reported that the US military used an Anthropic chatbot to help identify nearly 1,000 targets during the first 24 hours of its war with Iran. AI-assisted systems have also been used in other recent conflicts.

Separately, the Pentagon is revising its 2023 directive governing autonomous weapons, which requires appropriate levels of human judgement over the use of force.

President Donald Trump ordered an update of the directive in June, but the revised version had not been released by late September.